This article is part of our coverage of the latest in AI research.
There is increasing interest in using artificial intelligence to do scientific research. Several studies show AI systems, particularly large language models (LLM), autonomously do research and discover novel concepts.
However, a new paper by researchers at the Department of Psychology and Center for Brain Science at Harvard University argues that current AI systems are only solving the “easy problem” of science, and solving the “hard problem” of finding the problem itself is “beyond the capacities of current algorithms for scientific discovery.”
Once scientists formulate a function that they want to optimize, they need to find or collect a large dataset of annotated examples or verification mechanisms for the model’s output. This is the “easy problem” of science, which in all fairness is quite difficult.
Current AI systems are becoming increasingly more effective at this task, especially when it comes to exploring solution spaces that are too exhaustive for brute-force and simple search algorithms.
“What makes this problem ‘easy’ is not the form of the solution (which may require a great deal of engineering work) but rather the form of the problem,” the researchers write. “It is clear from the beginning what needs to be optimized, and what kinds of tools can be brought to bear on this problem. The engineering breakthrough comes from building much better versions of these tools.”
Algorithms that solve the easy problems of science are useful, even essential, to progress. We can see that in tools like AlphaFold 2 for protein folding. But these algorithms are provided a representation of the scientific problem that already includes the basic primitives needed for the final theory. Those primitives, along with the problem representation and the goal, are provided by the human scientist. This is the “hard problem.”
“In machine learning terms, these systems might be extremely good at interpolation, and they may become better at extrapolation to new data, but they will never automatically generate or choose to investigate new scientific problems,” the researchers write.
In the paper, the researchers go into detail about several historical case studies of scientific discoveries and their contemporary AI equivalents such as AI Feynman and AlphaFold 2. In each human discovery, the problem formulation and the solution space evolved together as the scientists carried out experiments, made new observations, rejected old hypotheses, reflected and discussed their findings, and looked at the problem in a new light.
In the case of AI tools, they are limited to searching within the confines of the space and constraints defined by the human. And setting the right constraints plays an important role in the success of the AI system.
Where do we go from here? “A research program for attacking the hard problem should begin with the cognitive science of science, focusing on the understudied subjective, creative aspects discussed above and how they interact with the objective aspects of problem solving,” the researchers write.
Only after understanding and formalizing what human scientists are doing with enough precision, we can try to leverage these insights to build scalable AI scientists. At first, these scientists will act like research assistants who need expert guidance through natural language instructions and demonstrations. Further improvement will require the study of all aspects of scientific discovery, including its social context.
“The growth of models beyond this requires the examination and emulation of the communal aspects of science and related cultural institutions,” the researchers write. “Lab meetings, conferences, and presentations and discussions are ultimately the place where judgements on the quality of a scientific problem are made.”